COMPARATIVE EVALUATION OF SHAP AND LIME FOR EXPLAINABLE TYPE 2 DIABETES PREDICTION MODELS

Abstract

Type 2 Diabetes Mellitus (T2DM) is a globally escalating chronic metabolic disorder affecting over 537 million adults worldwide as of 2021, with projections surpassing 780 million by 2045. Early and accurate prediction of T2DM through machine learning (ML) models offers significant clinical promise; however, the deployment of opaque, black-box predictive systems in healthcare raises substantial concerns regarding interpretability, accountability, and patient trust. Explainable Artificial Intelligence (XAI) methodologies have emerged as pivotal solutions to address these challenges. This paper presents a rigorous comparative evaluation of two leading XAI frameworks — SHapley Additive exPlanations (SHAP) and Local Interpretable Model-agnostic Explanations (LIME) — applied to multiple state-of-the-art ML classifiers for T2DM prediction, including Random Forest (RF), Gradient Boosting (XGBoost), LightGBM, Support Vector Machine (SVM), and Logistic Regression (LR). Experiments were conducted on the benchmark PIMA Indians Diabetes Dataset (PIDD) and an augmented clinical cohort. Key evaluation dimensions include predictive accuracy, explanation fidelity, computational efficiency, stability, and clinical interpretability. XGBoost with SHAP yielded the highest classification accuracy (AUC = 0.947), while LIME demonstrated superior explanation locality and model-agnostic adaptability. SHAP consistently produced more stable, globally coherent feature attributions compared to LIME's locally bounded, computationally intensive explanations. The findings provide actionable recommendations for clinical AI practitioners selecting XAI tools and highlight avenues for hybrid XAI architectures in diabetes management systems.